MSQA - Korean: leaderboard
Metric: F-score (%; harmonic mean of CO, the share of fully correct answers, and CGA, the correct share of attempted answers, over the 86 Korean-language questions; free-form answers to natively sourced culture-specific short questions asked in their own language, five runs per model, scored by a Gemini-3.1-Pro gold-answer containment judge). Source: arxiv.org. Saturation forecast: Around December 2026. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 73.9 |
| 2 | Claude Opus 4.7 | 62.8 |
| 3 | GPT-5.5 | 62.4 |
| 4 | GPT-5.4 | 59.7 |
| 5 | Claude Opus 4.6 | 56.4 |
| 6 | GPT-5.2 (High) | 46.6 |
| 7 | DeepSeek V4 Pro | 45.8 |
| 8 | Seed 2.1 Pro | 45.2 |
| 9 | Gemini 2.5 Flash | 41.2 |
| 10 | Seed 2.0 Pro (High) | 34 |
| 11 | Seed 2.0 Pro (Medium) | 32.6 |
| 12 | Seed 2.1 Turbo | 32 |
| 13 | Qwen 3.5 Plus (Thinking) | 31.7 |
| 14 | GLM-5 | 29.3 |
| 15 | Seed 2.0 Lite | 28.7 |
Interactive version: theaggregate.ai/benchmark?slug=msqa-korean · How It Works · Data refreshed daily, snapshot 2026-09-29.